This is a pre-print of a post that will be published on the collectors.com tech blog.
We recently went through one of our most epic product launches at Collectors: Customers are now able to submit cards for grading to PSA, add them to their collection on the new collectors.com site, and choose to have their valuable cards stored in an actual physical vault. Oh, and all of this will be accessible via a single login, their “Collectors ID”, which replaces the multiple logins customers previously had to manage across our different products and business units like PSA and PCGS. While all these new features are integrated to provide a seamless experience to our customers, we’re dealing with a significantly more complex architecture of multiple new systems, databases, and APIs on the backend. Naturally, as with any new product launch, our product managers were keen to get analytics about the use of these features from day 1: How many users actually converted to the new “Collectors ID”? How many items have customers submitted to the Vault? Who is using the new collectors.com “My Collection” feature?
In order to provide these kinds of data insights right from the go-live, we had to coordinate with several engineering teams to get our hands on the right data and integrate it into our data warehouse. In total, we ended up pulling data from systems sitting on top of four different production databases that were being launched at the same time. Considering how many different systems and databases we were working across, the integrations went pretty smoothly! Within a day of the go-live, I had produced a few dashboards with key metrics that the product and business stakeholders started using immediately to track uptake of the new services. The road to getting there wasn’t straightforward though and involved some amount of scrambling, knocking on different doors, and a few small surprises during the go-live. In this post, I will share some of my lessons learned from integrating with a new production system when you’re looking to provide analytics from the get-go.
I’m big on keeping running docs with notes from my conversations and findings when working on a project – I always say I outsource my brain into a Google doc. Keep a doc with (datestamped) notes and “to do” items for every piece of information you find, open questions, as well as a list of who’s responsible for what on the product, e.g. product managers, engineering leads, project managers, etc.
In addition to the running notes, connect with the business stakeholders (product managers, analysts…) early on to document a set of desired metrics along with their priorities and timelines: What do we need to know from day 1? What can wait until some time after the launch? This will also be helpful when exploring the new data models to determine what is actually being captured and what data points may not be available to calculate the required metrics.
If there are standing meetings for the engineering team that’s responsible for the database setup, I strongly recommend regularly sitting in on those meetings. Even if you don’t always understand everything that’s going on, it’s helpful to have the context of what the team is focusing on, and establish a relationship with them. As data engineers, we’re often pretty removed from our counterparts on the data producer side, but knowing the people on the team (and having them know you) can be helpful in working together more effectively.
Once you know who your engineering point of contact is, the first question you’ll want to ask is: How do we get access to the data? Assuming we’re talking about data that lives in a relational database, here’s a short check list of information you need to get from the engineering team that’s responsible for the database setup:
- Find out (and document) what cloud service the database is hosted in
- Will you get access to a production database or a read-replica? And what permissions will you get, read-only, or will you be able to create temp tables or views if they’re needed by any of the tools in your pipeline?
- Will there be dev and prod environments? What’s the timing for these being available?
- How do users and services authenticate against the database? Do we need personal and/or service accounts to log in?
- How will the logins will be shared? Will you need access to a shared password storage?
- Do you need an SSH tunnel setup to access the database from any of the tools in your data stack?
It’s best to try and get all these details ironed out as early as possible, since especially tasks like setting up SSH tunnels can take some time. Make sure you can access the database as early as possible to avoid surprises later on, even if there is no meaningful data in there yet.
Now that we’ve covered physical access to the data, let’s take a look at things to consider when you’re working with a new data model. I got looped into the production database design process early on and was able to provide input on the data modeling (see also: establishing a good connection with the upstream engineering teams! They’re your friends!). This ensured that the data would be suitable for our data extraction tool (Stitch) and contained all relevant data. Again, assuming you’re working with a relational database, here are some questions you’ll want to cover when talking about the data model:
- Where is the data model documentation and how is it being kept up to date?
- For any fields containing value sets, such as status codes, where are the corresponding descriptions stored? Will there be lookup tables in the database, or will these only be stored in code? The latter means you will need to be able to access the up-to-date list of lookups through your infrastructure, e.g. by querying an API (or simply reading the API documentation).
- Will there be JSON columns? What is the schema for those?
- What are the constraints on each table and column, e.g. foreign key relationships, NULL values, default values?
- For datetime fields, will they be stored with timezone (they should)?
Application and data flow
Perhaps most importantly, when trying to make sense of data coming from a production base, we need to understand what the flow of the application is: What workflows (user-created or automated) in the application modify the data in what way? This is absolutely crucial to handling the data correctly and drawing the right conclusions from it. For example:
- How and when is a record created, and what fields are populated through what input?
- What workflows cause records to be modified in what way? And what metadata is there to track modifications, e.g. a “last updated” timestamp?
- Will update timestamps for events such as status changes be tracked in separate fields? Or will there be kind of changelog table that captures these kinds of changes? This also trickles down into your data warehouse models, where you might need to start tracking status change dates right from the get-go.
- How are deletions being handled? Will there be “hard deletes”, i.e. the record is simply removed, or “soft deletes”, i.e. the record has a “is deleted” or “deleted timestamp” field. And, along the same lines, is there a data retention policy that means data will be dropped or archived after a certain amount of time?
- If the application is replacing a legacy application, will data be migrated? How do you recognize migrated data? Will there be any gaps or differences between migrated and newly create data?
- Will there be realistic dummy data (i.e. data that adheres to the constraints and workflows described above) to develop our data models and metrics against?
- Is there any chance of any test or dummy data getting into the production system? And if yes, how can we recognize and filter for it?
Ideally, your engineering and database admin teams will already have a “best practice” guide for designing new databases, which usually answers a lot of these questions. Otherwise, this might be a good time to start collecting these kinds of design decisions into a guide and encoding them in setup scripts where possible.
I hope that this post has provided you with a starting point for a checklist for your next production data integration. All the questions I’ve covered in the above paragraphs should be treated as conversation prompts to elicit existing design decisions, or help influence decisions that are yet to be made. There will likely be some oversights (I have yet to work with *the* perfect production database), but coming prepared with a plan may help you catch some of the biggest issues to getting a good data integration early on. And even with the best preparation, you can probably expect to make some tweaks after the application go-live to adjust to some last-minute database changes or correct some assumptions you’ve made about the data. Developing against an empty data model or even dummy data can be challenging, and you might not nail everything at first try.
One last thing to keep in mind: As data consumers, our downstream use case will most likely be of lower priority than getting the production system stood up – and that’s totally okay. While I would love for data to always be a first-class citizen, I believe it’s pretty obvious that producing a stable production system needs to take priority, and we just need to accept that resource constrained engineering teams may move slower on supporting a data integration. This is why you’ll want to get started early and get these kinds of tasks and questions on the engineering team’s radar as soon as possible.
CC-licensed photo by Ian Beckley: https://www.pexels.com/photo/top-view-photography-of-roads-2440013/